unknown

prediction: AGI will have a harness, but it will be just-in-time, self-assembled based on the current Task+context

today our human priors dominate the details of harness engineering. we inject our notions of good engineering practices into the harness design

but systems like alpha-zero have shown that this can be a local-minimum, models will get better than us at this eventually

it’s good to lean into more model autonomy while today leaning into good harness design using our priors

we want to get work done today!  that means lean into the frontier of using agents to do good work today which is practically a lot of context + harness eng

at the same time we can tinker at the frontier of the future by giving models way more autonomy

A prediction that AGI will use a just-in-time, self-assembled harness based on current task and context, contrasting with today's approaches.